基于SE-UNet的冬小麦种植区域提取方法

传统的小麦面积提取方法主要依靠人工野外调查,存在工作量大、效率低、成本高等问题,而遥感技术具有准确、快速和动态等优点,已成为作物面积提取的有效手段。本文以石家庄市正定县各镇的Landsat-8卫星遥感影像为训练数据,藁城区增村镇影像为测试数据,并分别选取分辨率8m的高分六号（GF-6）以及分辨率10m的哨兵二号（Sentinel-2）作为对比验证数据,提出了一种改进U-Net网络的冬小麦种植区域提取方法。首先,对Landsat-8遥感影像进行预处理,标注小麦区域制作标签集,其次,在U-Net网络中添加Squeeze and excitation(SE)注意力机制模块融入特征通道间信息,并利用Batch normalization(BN)层抑制过拟合问题;最后,经过Softmax分类器得到分类结果。选择SegNet、Deeplabv3+、U-Net作为对比模型,分别利用GF-6、Sentinel-2和Landsat-8 3种数据构建预测模型。结果表明,SE-UNet网络在基于Landsat-8数据预测模型下测试数据集表现最优,MPA和MIoU分别达到89.88%和81.44%。本方法可为大范围冬小麦种植区提取提供参考。

SE-UNet-Based Extraction of Winter Wheat Planting Areas

The traditional wheat area extraction methods mainly depend on artificial field investigation, which shows some disadvantages such as a big workload, low efficiency and high cost. Conversely, remote sensing technology has the advantages of high accuracy, rapid response and dynamic monitoring. It has become an effective measurement to extract crop areas. The Landsat-8 satellite remote sensing image of Zhengding County in Shijiazhuang was used as the training data, the image of Zengcun Town in Gaocheng District was used as test data. The GF-6 with resolution of 8m and Sentinel-2 with resolution of 10m were selected as comparative validation data. An improved U-Net was proposed to extract winter wheat planting areas. Landsat-8 was firstly preprocessed and the label set of wheat areas were marked and trained by using the U-Net network. The Squeeze and excitation (SE) attention mechanism module was introduced to better consider the information between feature channels, and the Batch normalization (BN) layer was used to suppress the over-fitting problem. The classification results were obtained through the Softmax classifier. SegNet, Deeplabv3+ and U-Net were selected as the comparison models and GF-6, Sentinel-2 and Landsat-8 data were used to construct the models, respectively. The results showed that the SE-UNet network performed best in the test data set based on Landsat-8 data prediction model, with the MPA and MIoU of 89.88% and 81.44%, respectively. This method can provide a reference for identifying large-scale winter wheat planting areas.